PyTorch Just Made Your Mac's AI Math Up to 2.4x Faster

PyTorch is the free software engine behind most modern AI. Nearly every chatbot, image generator, or voice tool you have used was trained or run with it. This update takes two common math routines — the ones that multiply a number and then add or divide it — and moves them onto the graphics chip inside Apple Silicon Macs. That chip is the part of your computer built for heavy number crunching. Doing this math there instead of on the main processor is like moving a pile of boxes from a narrow hallway to a loading dock.

The result, measured on an M4 Pro Mac, is up to 2.4 times faster for smaller jobs. Those are the everyday sizes most people actually work with. For very large jobs the speed is roughly the same as before, and for a few odd-shaped jobs the old method is still quicker. So this is a real improvement, not a miracle.

The team also fixed a hidden bug. When the software was told to save results in higher precision than it calculated with, it quietly did the whole calculation at that higher precision — slower — instead of just rounding at the end. Now it works like a calculator that keeps normal precision and rounds only the final answer. That also makes Mac results match what you'd get on other computers.

Why does this matter if you don't write code? Two reasons. First, AI tools running locally on a Mac — photo editors, transcription apps, chat assistants — get quicker and use less battery. Second, it means you depend less on the cloud. If your laptop can do more of the work itself, you send less data to someone else's server. The remaining slowdowns are flagged as future work.